elliothe / bfa Goto Github PK

When I was trying to perform the attacks on resnet18, I modified the model in BFA_imagenet.sh to resnet18, and got the following errors:

Traceback (most recent call last):
  File "main.py", line 860, in <module>
    main()
  File "main.py", line 445, in main
    random_attack=args.random_bfa)
  File "main.py", line 592, in perform_attack
    attack_log = attacker.progressive_bit_search(model, data, target)
  File "/home/yecheng/repo/BFA/attack/BFA.py", line 124, in progressive_bit_search
    key=operator.itemgetter(1))[0]
ValueError: max() arg is an empty sequence

I also tried the other Github repo for your previous paper, however, the issue persists.

May I get some help?

Hello, I tested BFA attack in this code but can't equal attack performance in vgg11-bn-quan, unlike resnet. (N_flips is 11 on paper, but 50 on test this code).

I think that the attack performance gap is caused by using different pre-trained weights.

So, plz uploaded your pre-trained vgg11-bn-quan weight?

I show that attack performance differs from BFA papers while attempting resnet44 and resnet56 BFA Attack. (paper resnet44 8bit is [6, 10, 11, 13, 22] but my resnet44 8bit is [57, 56, 59, 62, 60] and not drop acc 10%)

Therefore, I think BFA attack performance doesn't come out properly when the network gets deeper, what do you think?

Also, if possible, can you provide the weight parameter you used?

my resnet44 and resnet56 training script is this is this.
`
model=resnet56_quan
dataset=cifar10
epochs=160
train_batch_size=128
test_batch_size=128
optimizer=SGD
quan_bitwidth=8
seed=1
label_info=train

save_path=./save/${DATE}/${dataset}${model}${epochs}${optimizer}${label_info}${quan_bitwidth}${seed}
tb_path=${save_path}/tb_log #tensorboard log path

############### Neural network ############################
{
/usr/bin/env python ./main.py --dataset ${dataset}
--data_path ${data_path}
--arch ${model} --save_path ${save_path}
--quan_bitwidth ${quan_bitwidth}
--manualSeed ${seed}
--epochs ${epochs} --learning_rate 0.1
--optimizer ${optimizer}
--schedule 80 120 --gammas 0.1 0.1
--attack_sample_size ${train_batch_size}
--test_batch_size ${test_batch_size}
--workers 4 --ngpu 1 --gpu_id 0
--print_freq 100 --decay 0.0003 --momentum 0.9
} &

model=resnet44_quan
dataset=cifar10
epochs=160
train_batch_size=128
test_batch_size=128
optimizer=SGD
quan_bitwidth=8
seed=1
label_info=train

save_path=./save/${DATE}/${dataset}${model}${epochs}${optimizer}${label_info}${quan_bitwidth}${seed}
tb_path=${save_path}/tb_log #tensorboard log path

############### Neural network ############################
{
/usr/bin/env python ./main.py --dataset ${dataset}
--data_path ${data_path}
--arch ${model} --save_path ${save_path}
--quan_bitwidth ${quan_bitwidth}
--manualSeed ${seed}
--epochs ${epochs} --learning_rate 0.1
--optimizer ${optimizer}
--schedule 80 120 --gammas 0.1 0.1
--attack_sample_size ${train_batch_size}
--test_batch_size ${test_batch_size}
--workers 4 --ngpu 1 --gpu_id 0
--print_freq 100 --decay 0.0003 --momentum 0.9
} &

File "main.py", line 860, in
main()
File "main.py", line 445, in main
random_attack=args.random_bfa)
File "main.py", line 572, in perform_attack
attacker.criterion, log, summary_output=True)
File "main.py", line 770, in validate
prec1, prec5 = accuracy(output.data, target, topk=(1, 5))
File "main.py", line 835, in accuracy
correct_k = correct[:k].view(-1).float().sum(0)
RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead.

The bin_resnet_cifar code present in the corresponding model folder seems to be quantization resnet rather than bin_resnet_cifar.

Can i refer to bin_vgg that convert to NN from quantized to binarized ?